WO2017094121A1

WO2017094121A1 - Voice recognition device, voice emphasis device, voice recognition method, voice emphasis method, and navigation system

Info

Publication number: WO2017094121A1
Application number: PCT/JP2015/083768
Authority: WO
Inventors: 勇気太刀岡
Original assignee: 三菱電機株式会社
Priority date: 2015-12-01
Filing date: 2015-12-01
Publication date: 2017-06-08
Also published as: JP6289774B2; CN108292501A; JPWO2017094121A1; KR20180063341A; US20180350358A1; TW201721631A; DE112015007163T5; KR102015742B1; DE112015007163B4

Abstract

This invention is provided with: a plurality of noise suppression units (3) for performing noise suppression processes of mutually different methods on inputted noise audio data; a voice recognition unit (4) for performing voice recognition on voice data in which noise signals are suppressed; a prediction unit (2) for predicting, from acoustic feature quantities of inputted noise audio data, the voice recognition rate obtained when a noise suppression process is performed on the noise audio data by each of the plurality of noise suppression units (3); and a suppression method selection unit (2) for selecting, on the basis of the predicted voice recognition rates, the noise suppression unit (3) performing a noise suppression process on the noise audio data from the plurality of noise suppression units.

Description

Speech recognition device, speech enhancement device, speech recognition method, speech enhancement method, and navigation system

The present invention relates to a speech recognition technology and a speech enhancement technology, and particularly to a technology corresponding to use under various noise environments.

When performing speech recognition using speech with superimposed noise, it is common to perform processing for suppressing superimposed noise (hereinafter referred to as noise suppression processing) before performing speech recognition processing. Depending on the characteristics of the noise suppression processing, there are noises that are effective and noises that are not effective for the noise suppression processing. For example, when the noise suppression process is a spectrum removal process that is strong against stationary noise, the removal process against unsteady noise becomes weak. On the other hand, when the noise suppression process is a process that has high followability to non-stationary noise, the process is low in followability to steady noise. Conventionally, integration of speech recognition results or selection of speech recognition results has been used as a method for solving such problems.

In the conventional method, when a sound with superimposed noise is input, noise is generated by two noise suppression units that perform, for example, suppression processing with high followability to stationary noise and suppression processing with high followability to non-stationary noise. The two voices are acquired by suppressing the voice, and the voices are recognized by the two voice recognition units with respect to the obtained two voices. The two speech recognition results obtained by speech recognition are integrated using a speech combining method such as ROVER (Recognition Output Voting Error Reduction), or the speech recognition result with the highest likelihood is selected from the two speech recognition results, The integrated or selected speech recognition result is output. However, the conventional method has a problem that the degree of improvement in recognition accuracy is large, but processing for speech recognition increases.

As a technique for solving this problem, for example, Patent Document 1 discloses a speech recognition device that calculates the likelihood of each acoustic feature parameter of input noise for each probability speech model and selects a sound probability acoustic model from the likelihood. ing. Japanese Patent Laid-Open No. 2004-228688 discloses a method in which noise is removed from an input target signal, preprocessing for extracting feature amount data representing the characteristics of the target signal is performed, and then the target signal is determined according to the shape of the clustering map of the competitive neural network. Is classified into a plurality of categories, and a signal identification device that automatically selects processing contents is disclosed.

JP 2000-194392 A JP 2005-115569 A

However, since the technique disclosed in Patent Document 1 described above uses the likelihood of each acoustic feature parameter of the input noise for each probability speech model, a noise suppression process that can obtain a good speech recognition rate or acoustic index is selected. There was a problem that it might not be. In the technique disclosed in Patent Document 2, although the target signal is clustered, clustering associated with the speech recognition rate or the acoustic index is not performed. There was a problem that the obtained noise suppression processing might not be selected. In addition, the above two methods share the need for noise-suppressed speech for performance prediction, so all candidate noise suppression processing must be performed once during both learning and use. There was a problem.

The present invention has been made in order to solve the above-described problems, and provides a good speech recognition rate or acoustic index only from noise speech data without performing noise suppression processing at the time of use in order to select a noise suppression method. It is an object to select a noise suppression process that provides high accuracy.

A speech recognition apparatus according to the present invention includes a plurality of noise suppression units that perform noise suppression processing using different methods on input noise speech data, and speech recognition of speech data in which noise signals are suppressed by the noise suppression unit. A speech recognition unit that performs a noise recognition process, and a prediction unit that predicts a speech recognition rate obtained when noise suppression processing is performed on each of the noise speech data by a plurality of noise suppression units, based on the acoustic feature amount of the input noise speech data, And a suppression method selection unit that selects a noise suppression unit that performs noise suppression processing on noise speech data from a plurality of noise suppression units based on the speech recognition rate predicted by the prediction unit.

According to the present invention, it is possible to select a noise suppression process that provides a good speech recognition rate or acoustic index without performing a noise suppression process to select a noise suppression method.

1 is a block diagram showing a configuration of a speech recognition apparatus according to Embodiment 1. FIG. 2A and 2B are diagrams showing a hardware configuration of the speech recognition apparatus according to the first embodiment. 3 is a flowchart showing an operation of the speech recognition apparatus according to the first embodiment. 4 is a block diagram illustrating a configuration of a speech recognition apparatus according to Embodiment 2. FIG. 6 is a flowchart showing the operation of the speech recognition apparatus according to the second embodiment. FIG. 6 is a block diagram illustrating a configuration of a speech recognition apparatus according to a third embodiment. It is a figure which shows the structural example of the recognition rate database of the speech recognition apparatus which concerns on Embodiment 3. FIG. 10 is a flowchart showing the operation of the speech recognition apparatus according to the third embodiment. FIG. 10 is a block diagram illustrating a configuration of a speech enhancement device according to a fourth embodiment. 10 is a flowchart illustrating an operation of the speech enhancement device according to the fourth embodiment. FIG. 10 is a functional block diagram showing a configuration of a navigation system according to a fifth embodiment.

Hereinafter, in order to explain the present invention in more detail, modes for carrying out the present invention will be described with reference to the accompanying drawings.
Embodiment 1 FIG.
First, FIG. 1 is a block diagram showing the configuration of the speech recognition apparatus 100 according to the first embodiment.
The speech recognition apparatus 100 includes a first prediction unit 1, a suppression method selection unit 2, a noise suppression unit 3, and a speech recognition unit 4.
The first prediction unit 1 is composed of a regressor. For example, a neural network (hereinafter referred to as NN) is constructed and applied as the regressor. In the construction of the NN, an acoustic feature that is generally used, such as Mel-frequency Cepstral Coefficient (MFCC) or filter bank feature, is used to directly calculate a speech recognition rate of 0 to 1 as a regressor. Is constructed using, for example, the error back-propagation method. The error back-propagation method is a learning method in which, when given learning data is given, the coupling load / bias between the layers is corrected so that the error between the learning data and the output of the NN becomes small. The first predicting unit 1 predicts the speech recognition rate of the input acoustic feature amount by, for example, the NN having the input as the acoustic feature amount and the output as the speech recognition rate.

The suppression method selection unit 2 refers to the speech recognition rate predicted by the first prediction unit 1, and selects the noise suppression unit 3 that performs noise suppression from the plurality of

noise suppression units

3a, 3b, and 3c. The suppression method selection unit 2 outputs a control instruction to perform noise suppression processing on the selected noise suppression unit 3. The noise suppression unit 3 includes a plurality of

noise suppression units

3a, 3b, and 3c, and each

noise suppression unit

3a, 3b, and 3c performs different noise suppression processing on the input noise voice data. As different noise suppression processing, for example, an adaptive filter method using a spectrum removal method (SS), a learning identification method (Normalized Least Mean Square Algorithm; NLMS algorithm), or a method using NN such as Denoising auto encoder can be applied. It is. Also, which of the

noise suppression units

3a, 3b, and 3c performs the noise suppression processing is determined based on a control instruction input from the suppression method selection unit 2. In the example of FIG. 1, an example in which the three

noise suppression units

3a, 3b, and 3c are configured is shown, but the number of components is not limited to three and can be changed as appropriate.

The voice recognition unit 4 performs voice recognition on the voice data in which the noise signal is suppressed by the noise suppression unit 3, and outputs a voice recognition result. For speech recognition, speech recognition processing is performed using, for example, an acoustic model based on Gaussian mix model or Deep neural network and a language model based on n-gram. Note that since the voice recognition process can be configured by applying a known technique, a detailed description thereof will be omitted.

The first prediction unit 1, the suppression method selection unit 2, the noise suppression unit 3, and the speech recognition unit 4 of the speech recognition apparatus 100 are realized by a processing circuit. The processing circuit may be dedicated hardware or may be a CPU (Central Processing Unit) that executes a program stored in a memory, a processing device, a processor, and the like.
FIG. 2A shows a hardware configuration of the speech recognition apparatus 100 according to Embodiment 1, and shows a block diagram when the processing circuit is executed by hardware. As shown in FIG. 2A, when the processing circuit 101 is dedicated hardware, each function of the first prediction unit 1, the suppression method selection unit 2, the noise suppression unit 3, and the speech recognition unit 4 is realized by the processing circuit. Alternatively, the functions of the respective units may be integrated and realized by a processing circuit.

FIG. 2B shows a hardware configuration of the speech recognition apparatus 100 according to Embodiment 1, and shows a block diagram when the processing circuit is executed by software.
As shown in FIG. 2B, when the processing circuit is the processor 102, each function of the first prediction unit 1, the suppression method selection unit 2, the noise suppression unit 3, and the speech recognition unit 4 is software, firmware, or software. Realized by combination with firmware. Software and firmware are described as programs and stored in the memory 103. The processor 102 reads out and executes the program stored in the memory 103, thereby executing the function of each unit. Here, the memory 103 corresponds to, for example, a nonvolatile or volatile semiconductor memory such as a RAM, a ROM, and a flash memory, a magnetic disk, and an optical disk.

As described above, the processing circuit can realize the above-described functions by hardware, software, firmware, or a combination thereof.

Next, detailed configurations of the first prediction unit 1 and the suppression method selection unit 2 will be described.
First, the 1st prediction part 1 to which the regressor is applied is comprised by NN which uses an acoustic feature-value as an input and uses an output as a speech recognition rate. When the acoustic feature quantity is input for each frame of the short-time Fourier transform, the first prediction unit 1 predicts the speech recognition rate by each of the

noise suppression units

3a, 3b, and 3c by NN. That is, the first predicting unit 1 calculates a speech recognition rate for each frame of the acoustic feature amount when different noise suppression processes are applied. The suppression method selection unit 2 refers to the speech recognition rate when the

noise suppression units

3a, 3b, and 3c calculated by the first prediction unit 1 are applied, and noise suppression that leads to a speech recognition result with the highest speech recognition rate The unit 3 is selected, and a control instruction is output to the selected noise suppression unit 3.

FIG. 3 is a flowchart showing the operation of the speech recognition apparatus 100 according to the first embodiment.
It is assumed that the voice recognition apparatus 100 receives noise voice data and an acoustic feature amount of the noise voice data via an external microphone or the like. It is assumed that the acoustic feature amount of the noise voice data is calculated by an external feature amount calculation unit.
When the noise voice data and the acoustic feature quantity of the noise voice data are input (step ST1), the first predicting unit 1 performs each noise by NN in a frame unit of short-time Fourier transform of the input acoustic feature quantity. A speech recognition rate is predicted when noise suppression processing is performed by the

suppression units

3a, 3b, and 3c (step ST2). Note that the process of step ST2 is repeatedly performed for a plurality of set frames. When the first prediction unit 1 obtains the average, maximum value, or minimum value of the speech recognition rate predicted for each frame and a plurality of frames in step ST2, and performs processing in each

noise suppression unit

3a, 3b, 3c Each prediction recognition rate is calculated (step ST3). The first prediction unit 1 associates the calculated prediction recognition rate with each

noise suppression unit

3a, 3b, 3c and outputs it to the suppression method selection unit 2 (step ST4).

The suppression technique selection unit 2 refers to the prediction recognition rate output in step ST4, selects the noise suppression unit 3 that exhibits the highest prediction recognition rate, and performs noise suppression processing on the selected noise suppression unit 3. A control instruction is output to (step ST5). The noise suppression unit 3 to which the control instruction is input in step ST5 performs a process of suppressing the noise signal with respect to the actual noise voice data input in step ST1 (step ST6). The voice recognition unit 4 performs voice recognition on the voice data in which the noise signal is suppressed in step ST6, acquires a voice recognition result, and outputs it (step ST7). Thereafter, the flowchart returns to the process of step ST1 and repeats the process described above.

As described above, according to the first embodiment, the first predicting unit 1 configured by the regressor, configured by the NN having the acoustic feature quantity as input and the output as the speech recognition rate, With reference to the speech recognition rate predicted by the prediction unit 1, the noise suppression unit 3 that leads the speech recognition result with the highest speech recognition rate is selected from the plurality of noise suppression units 3, and a control instruction is given to the selected noise suppression unit 3 And a noise suppression unit 3 that performs noise suppression processing of noise speech data based on a control instruction of the suppression method selection unit 2, and a suppression method selection unit 2 that outputs a plurality of processing units to which a plurality of noise suppression methods are applied. The voice recognition unit 4 that performs voice recognition of the voice data subjected to the noise suppression processing is provided, so that the noise suppression can be performed without increasing the processing amount of voice recognition and selecting a noise suppression method. Effective noise suppression without processing It can be selected.
For example, in the conventional technology, when there are three candidate noise suppression methods, the noise suppression processing is performed by all three methods and the best noise suppression processing is selected based on the result. According to the first aspect, even when there are three candidate noise suppression methods, a method that will have the best performance can be predicted in advance. Therefore, noise suppression processing is performed by performing noise suppression processing only with the selected method. There is an advantage that the amount of calculation for processing can be reduced.

Embodiment 2. FIG.
In the first embodiment described above, the configuration has been shown in which the noise suppression unit 3 that leads the speech recognition result having a high speech recognition rate is selected using the regressor. However, in the second embodiment, the speech recognition rate is determined using the discriminator. The structure which selects the noise suppression part 3 which guides a high speech recognition result is shown.
FIG. 4 is a block diagram illustrating a configuration of the speech recognition apparatus 100a according to the second embodiment.
The speech recognition apparatus 100a according to Embodiment 2 replaces the first prediction unit 1 and the suppression method selection unit 2 of the speech recognition apparatus 100 described in Embodiment 1, with a second prediction unit 1a and a suppression method selection unit 2a. Is provided. In the following description, the same or corresponding parts as the components of the speech recognition apparatus 100 according to the first embodiment are denoted by the same reference numerals as those used in the first embodiment, and the description thereof is omitted or simplified. .

The second prediction unit 1a is composed of a discriminator. As the discriminator, for example, NN is constructed and applied. In the construction of NN, a generally used acoustic feature such as MFCC or filter bank feature is used, and classification processing such as 2-class classification or multi-class classification is performed as a discriminator, and the suppression method with the highest recognition rate The NN that selects the identifiers of the two is constructed using the error back-propagation method. The second prediction unit 1a performs, for example, an acoustic feature value as an input, performs a 2-class or multi-class classification with a final output layer as a softmax layer, and uses a suppression method ID that leads to a speech recognition result with the highest speech recognition rate It is composed of NN as (identification). For NN teacher data, only the suppression method that leads to the speech recognition result with the highest speech recognition rate is set to “1”, the other methods are set to “0”, the recognition rate is multiplied by Sigmoid, etc. Weighted data (Sigmoid ((recognition rate of the system— (max (recognition rate) −min (recognition rate) / 2)) / σ) can be used, where σ is a scaling factor.
Of course, it is possible to use other classifiers such as SVM (support vector machine).

The suppression method selection unit 2a refers to the suppression method ID predicted by the second prediction unit 1a, and selects the noise suppression unit 3 that performs noise suppression from the plurality of

noise suppression units

3a, 3b, and 3c. As in the first embodiment, the noise suppression unit 3 can employ a spectrum removal method (SS), an adaptive filter method, a method using NN, and the like. The suppression method selection unit 2a outputs a control instruction to the selected noise suppression unit 3 so as to perform noise suppression processing.

Next, the operation of the speech recognition apparatus 100a will be described.
FIG. 5 is a flowchart showing the operation of the speech recognition apparatus 100a according to the second embodiment. In the following, the same steps as those of the speech recognition apparatus 100 according to Embodiment 1 are denoted by the same reference numerals as those used in FIG. 3, and the description thereof is omitted or simplified.
It is assumed that noise voice data and an acoustic feature amount of the noise voice data are input to the voice recognition device 100a via an external microphone or the like.
When the noise voice data and the acoustic feature amount of the noise voice data are input (step ST1), the second predicting unit 1a is the most speech by the NN in a short-time Fourier transform frame unit of the input acoustic feature amount. A suppression method ID of a noise suppression method that leads to a speech recognition result with a high recognition rate is predicted (step ST11).

The second prediction unit 1a obtains the mode or average value of the suppression method ID predicted in units of frames in step ST11, and acquires the mode or average suppression method ID as the prediction suppression method ID (step S11). ST12). The suppression method selection unit 2a refers to the prediction suppression method ID acquired in step ST12, selects the noise suppression unit 3 corresponding to the acquired prediction suppression method ID, and performs noise suppression processing on the selected noise suppression unit 3. A control instruction is output so as to be performed (step ST13). Thereafter, the same processing as that in step ST6 and step ST7 described in the first embodiment is performed.

As described above, according to the second embodiment, a discriminator is applied, and an acoustic feature quantity is input, and an output is composed of NNs that are IDs of suppression methods that lead to a speech recognition result with the highest speech recognition rate. The second prediction unit 1a and the noise suppression unit 3 that leads the speech recognition result with the highest speech recognition rate from the plurality of noise suppression units 3 with reference to the suppression method ID predicted by the second prediction unit 1a are selected. And a suppression method selection unit 2a that outputs a control instruction to the selected noise suppression unit 3, and a plurality of processing units corresponding to each of the plurality of noise suppression processes, and noise based on the control instruction of the suppression method selection unit 2a Since it is configured to include the noise suppression unit 3 that performs noise suppression of speech data and the speech recognition unit 4 that performs speech recognition of speech data that has undergone noise suppression processing, the processing amount of speech recognition is not increased. And noise suppression techniques Without performing the noise suppression process for-option, it is possible to select a valid noise suppression techniques.

Embodiment 3 FIG.
In the first and second embodiments described above, the acoustic feature amount is input to the first prediction unit 1 or the second prediction unit 1a for each short-time Fourier transform frame, and the speech recognition rate or suppression is input for each input frame. The structure which estimates method ID was shown. On the other hand, in the third embodiment, an utterance closest to the acoustic feature amount of the noise speech data actually input to the speech recognition device is selected from previously learned data using the acoustic feature amount of the utterance unit. 1 shows a configuration for selecting a noise suppression unit based on a speech recognition rate of a selected utterance.

FIG. 6 is a block diagram illustrating a configuration of the speech recognition apparatus 100b according to the third embodiment.
The speech recognition apparatus 100b according to the third embodiment replaces the first prediction unit 1 and the suppression technique selection unit 2 of the speech recognition apparatus 100 described in the first embodiment, and includes a feature amount calculation unit 5 and a similarity calculation unit 6. The third prediction unit 1c and the suppression method selection unit 2b including the recognition rate database 7 are provided.
In the following description, the same or corresponding parts as the components of the speech recognition apparatus 100 according to the first embodiment are denoted by the same reference numerals as those used in the first embodiment, and the description thereof is omitted or simplified. .

The feature quantity calculation unit 5 constituting the third prediction unit 1c calculates an acoustic feature quantity in units of utterances from the input noise voice data. The details of the method for calculating the acoustic feature amount for each utterance will be described later. The similarity calculation unit 6 refers to the recognition rate database 7, collates the acoustic feature amount of the utterance unit calculated by the feature amount calculation unit 5 with the acoustic feature amount stored in the recognition rate database 7, and determines the acoustic feature amount. The similarity is calculated. The similarity calculation unit 6 acquires a set of speech recognition rates when noise suppression is performed by the

noise suppression units

3a, 3b, and 3c associated with the acoustic feature amount having the highest similarity among the calculated similarities. And output to the suppression method selection unit 2b. Examples of the speech recognition rate pairs are “speech recognition rate _1-1 , speech recognition rate _1-2 , speech recognition rate _1-3 ” and “speech recognition rate _2-1 , speech recognition rate _2-2 , speech recognition rate”. _2-3 ". The suppression method selection unit 2b refers to the set of speech recognition rates input from the similarity calculation unit 6, and selects the noise suppression unit 3 that performs noise suppression from the plurality of

noise suppression units

3a, 3b, and 3c.

The recognition rate database 7 is a storage area in which acoustic feature amounts of a plurality of learning data and speech recognition rates when the

noise feature units

3a, 3b, and 3c suppress noise are associated and stored. .
FIG. 7 is a diagram illustrating a configuration example of the recognition rate database 7 of the speech recognition apparatus 100b according to the third embodiment.
The recognition rate database 7 includes the acoustic feature amount of the learning data, and the speech in which each learning data is subjected to noise suppression processing by each noise suppression unit (first, second, and third noise suppression units in the example of FIG. 7). The voice recognition rate of the data is stored in association with each other. In FIG. 7, for example, the speech recognition rate of the speech data in which the first noise suppression unit has performed the noise suppression processing with respect to the learning data of the first acoustic feature amount V ^(r1) is 80%, and the second The voice recognition rate of the voice data subjected to the noise suppression processing by the noise suppression unit is 75%, and the voice recognition rate of the voice data subjected to the noise suppression processing by the third noise suppression unit is 78%. ing. Note that the recognition rate database 7 may be configured to cluster the learning data, store the recognition rate of the clustered learning data and the acoustic feature amount in association with each other, and store the data amount while suppressing the data amount.

Next, the details of the calculation of the acoustic feature amount for each utterance by the feature amount calculation unit 5 will be described.
As the acoustic feature quantity of the utterance unit, an average vector of acoustic feature quantities, an average likelihood vector by Universal background model (UBM), an i-vector, or the like can be applied. The feature amount calculation unit 5 calculates the above-described acoustic feature amount in units of utterances for each of the recognition target noise sound data. For example, when an i-vector is applied as an acoustic feature, a Gaussian mixture model (GMM) is applied to the utterance r, and the obtained supervector V ^(r) is obtained in advance as a UBM supervector. Factorization is performed based on the following equation (1) by a matrix T composed of basis vectors extending v and all rank planes of low rank.
V ^(r) = v + Tw ^(r) (1)
W ^(r) obtained by the above equation (1 ⁾ is an i-vector.

The similarity between acoustic features of the utterance unit, as shown in the following equation (2), measure using a Euclid distance or cosine similarity closest to the current evaluation data r _e from in the training data r _t to select a speech _r't. When the similarity is represented by sim, an utterance represented by the following formula (3) is selected.

If the word error rate W _tr (i, r _t ) obtained using the i-th noise suppression unit 3 and the speech recognition unit 4 is obtained in advance for the learning data r _t , it is optimal for re The system i ′ is selected based on the recognition performance as shown in the following equation (4).

In the above description, the case where there are two noise suppression methods has been described as an example, but the present invention can also be applied to a case where there are three or more noise suppression methods.

Next, the operation of the speech recognition apparatus 100b will be described.
FIG. 8 is a flowchart showing the operation of the speech recognition apparatus 100b according to the third embodiment. In the following, the same steps as those of the speech recognition apparatus 100 according to Embodiment 1 are denoted by the same reference numerals as those used in FIG. 3, and the description thereof is omitted or simplified.
It is assumed that noise voice data is input to the voice recognition device 100b via an external microphone or the like.
When noise voice data is input (step ST21), the feature amount calculation unit 5 calculates an acoustic feature amount from the input noise voice data (step ST22). The similarity calculation unit 6 compares the acoustic feature quantity calculated in step ST22 with the acoustic feature quantity of the learning data stored in the recognition rate database 7, and calculates the similarity degree (step ST23). The similarity calculation unit 6 selects the acoustic feature amount indicating the highest similarity among the similarity of the acoustic feature amount calculated in step ST23, and is associated with the selected acoustic feature amount with reference to the recognition rate database 7. A set of recognition rates is acquired (step ST24). In step ST24, when the Euclid distance is used as the similarity between the acoustic feature quantities, a pair of recognition rates with the shortest distance is acquired.

The suppression technique selection unit 2b selects the noise suppression unit 3 that exhibits the highest recognition rate from the recognition rate group acquired in step ST24, and performs noise suppression processing on the selected noise suppression unit 3. A control instruction is output (step ST25). Thereafter, the same processing as in steps ST6 and ST7 described above is performed.

As described above, according to the third embodiment, the calculated acoustic feature amount and the acoustic feature of the learning data are referred to by referring to the feature amount calculation unit 5 that calculates the acoustic feature amount from the noise voice data and the recognition rate database 7. The similarity calculation unit 6 that calculates the similarity to the amount and acquires the speech recognition rate pair associated with the acoustic feature amount indicating the highest similarity, and the highest among the acquired speech recognition rate pairs Since it comprises the suppression method selection part 2b which selects the noise suppression part 3 which shows a speech recognition rate, the speech recognition performance can be predicted in speech units, and the speech recognition performance is highly predicted and fixed. By using the dimension feature amount, there is an effect that the similarity can be easily calculated.

In the above-described third embodiment, the configuration in which the speech recognition apparatus 100b includes the recognition rate database 7 has been described. However, the similarity calculation unit 6 refers to an external database and calculates the similarity with the acoustic feature amount. You may comprise so that acquisition of a recognition rate may be performed.

Note that in Embodiment 3 described above, a delay occurs when speech recognition is performed in units of utterances, but if the delay cannot be allowed, the acoustic feature value is calculated using the utterances for the first few seconds after the start of the utterance. You may comprise so that it may refer. Further, in the case where the environment with the utterance made before the utterance subject to the speech recognition does not change, the speech recognition is performed using the selection result of the noise suppression unit 3 in the previous utterance. Also good.

Embodiment 4 FIG.
In the third embodiment described above, the configuration in which the noise suppression method is selected with reference to the recognition rate database 7 in which the acoustic feature amount of the learning data is associated with the speech recognition rate has been described. The structure which selects the noise suppression method with reference to the acoustic parameter | index database which matched the acoustic feature-value and acoustic parameter | index of data is shown.
FIG. 9 is a block diagram showing the configuration of the speech enhancement apparatus 200 according to the fourth embodiment.
The speech enhancement apparatus 200 according to the fourth embodiment includes the third prediction unit 1c and the suppression provided with the feature amount calculation unit 5, the similarity calculation unit 6, and the recognition rate database 7 of the speech recognition apparatus 100b illustrated in the third embodiment. Instead of the method selection unit 2b, a fourth prediction unit 1d and a suppression method selection unit 2c each including a feature amount calculation unit 5, a similarity calculation unit 6a, and an acoustic index database 8 are provided. Further, the voice recognition unit 4 is not provided.
In the following, the same or corresponding parts as those of the speech recognition apparatus 100b according to the third embodiment are denoted by the same reference numerals as those used in the third embodiment, and the description thereof is omitted or simplified. .

The acoustic index database 8 is a storage area in which acoustic feature quantities of a plurality of learning data are stored in association with acoustic indices when the respective learning data are noise-suppressed by the

noise suppression units

3a, 3b, and 3c. Here, the acoustic index is PESQ or SNR / SDR calculated from the emphasized speech in which the noise is suppressed and the noise speech before the noise is suppressed. Note that the acoustic index database 8 may be configured to cluster the learning data, store the acoustic index of the clustered learning data and the acoustic feature amount in association with each other, and store the data while suppressing the data amount.

The similarity calculation unit 6 a refers to the acoustic index database 8, collates the acoustic feature amount of the utterance unit calculated by the feature amount calculation unit 5 with the acoustic feature amount stored in the acoustic index database 8, and determines the acoustic feature amount The similarity is calculated. The similarity calculation unit 6a acquires a set of acoustic indices associated with the acoustic feature quantity having the highest similarity among the calculated similarities, and outputs it to the suppression method selection unit 2c. Examples of the set of acoustic indicators include “PESQ _1-1 , PESQ _1-2 , PESQ _1-3 ” and “PESQ _2-1 , PESQ _2-2 , PESQ _2-3 ” and the like.
The suppression method selection unit 2c refers to the set of acoustic indices input from the similarity calculation unit 6a, and selects the noise suppression unit 3 that performs noise suppression from the plurality of

noise suppression units

3a, 3b, and 3c.

Next, the operation of the speech enhancement apparatus 200 will be described.
FIG. 10 is a flowchart showing the operation of the speech enhancement apparatus 200 according to the fourth embodiment. It is assumed that noise speech data is input to the speech enhancement device 200 via an external microphone or the like.
When noise voice data is input (step ST31), the feature quantity calculation unit 5 calculates an acoustic feature quantity from the input noise voice data (step ST32). The similarity calculation unit 6a compares the acoustic feature quantity calculated in step ST32 with the acoustic feature quantity of the learning data stored in the acoustic index database 8, and calculates the similarity degree (step ST33). The similarity calculation unit 6a selects an acoustic feature amount that indicates the highest similarity among the similarities of the acoustic feature amounts calculated in step ST33, and acquires a set of acoustic indicators associated with the selected acoustic feature amount. (Step ST34).

The suppression technique selection unit 2c selects the noise suppression unit 3 that indicates the highest acoustic index among the set of acoustic indexes acquired in step ST34, and performs noise suppression processing on the selected noise suppression unit 3. A control instruction is output (step ST35). The noise suppression unit 3 to which the control instruction is input in step ST35 performs processing for suppressing the noise signal on the actual noise speech data input in step ST31 to acquire and output the emphasized speech (step ST36). . Thereafter, the flowchart returns to the process of step ST31 and repeats the above-described process.

As described above, according to the fourth embodiment, the calculated acoustic feature quantity and the acoustic feature of the learning data are referred to by referring to the feature quantity calculation unit 5 that calculates the acoustic feature quantity from the noise voice data and the acoustic index database 8. A similarity calculation unit 6a for calculating a similarity to a quantity and acquiring a set of acoustic indices associated with an acoustic feature quantity indicating the highest similarity, and the highest acoustic index among the acquired sets of acoustic indices And the suppression method selection unit 2c that selects the noise suppression unit 3 indicating the voice recognition performance can be predicted in units of utterances, the speech recognition performance is highly predicted, and fixed-dimension features By using the quantity, there is an effect that the calculation of similarity becomes easy.

In the fourth embodiment described above, the configuration in which the speech enhancement apparatus 200 includes the acoustic index database 8 has been described. However, the similarity calculation unit 6a refers to an external database and calculates the similarity with the acoustic feature amount. You may comprise so that an acoustic parameter | index may be acquired.

In Embodiment 4 described above, a delay occurs when speech recognition is performed in units of utterances, but if the delay cannot be tolerated, the acoustic feature value is calculated using the utterances for the first few seconds after the start of the utterance. You may comprise so that it may refer. Further, when the environment with the utterance performed before the utterance that is the target of the enhanced speech acquisition does not change, the enhanced speech is acquired using the selection result of the noise suppression unit 3 in the previous utterance. It may be configured.

Embodiment 5 FIG.
The

speech recognition devices

100, 100a, and 100b of Embodiment 1-3 and the speech enhancement device 200 of Embodiment 4 described above are applied to, for example, a navigation system, a telephone-compatible system, and an elevator that have a voice call function. Can do. In the fifth embodiment, the case where the speech recognition apparatus of the first embodiment is applied to a navigation system will be described.
FIG. 11 is a functional block diagram showing a configuration of the navigation system 300 according to the fifth embodiment.
The navigation system 300 is a device that is mounted on a vehicle and performs route guidance to a destination, for example, and includes an information acquisition device 301, a control device 302, an output device 303, an input device 304, a voice recognition device 100, a map database 305, A route calculation device 306 and a route guidance device 307 are provided. The operation of each device of the navigation system 300 is centrally controlled by the control device 302.

The information acquisition device 301 includes, for example, current position detection means, wireless communication means, surrounding information detection means, and the like, and acquires information detected by the current position of the host vehicle, the surroundings of the host vehicle, and other vehicles. The output device 303 includes, for example, a display unit, a display control unit, a voice output unit, a voice control unit, and the like, and notifies the user of information. The input device 304 is realized by voice input means such as a microphone and operation input means such as buttons and a touch panel, and receives information input from the user. The speech recognition device 100 is a speech recognition device having the configuration and functions described in the first embodiment, performs speech recognition on noise speech data input via the input device 304, and acquires a speech recognition result. And output to the control device 302.

The map database 305 is a storage area for storing map data, and is realized as a storage device such as an HDD (Hard Disk Drive) or a RAM (Random Access Memory). The route calculation device 306 uses the current position of the vehicle acquired by the information acquisition device 301 as a departure point, uses the voice recognition result of the voice recognition device 100 as a destination, and based on the map data stored in the map database 305. The route from to the destination is calculated. The route guidance device 307 guides the host vehicle according to the route calculated by the route calculation device 306.

When the noise speech data including the user's utterance is input from the microphone constituting the input device 304, the speech recognition device 100 performs the processing shown in the flowchart of FIG. 3 on the noise speech data. To obtain a speech recognition result. Based on the information input from the control device 302 and the information acquisition device 301, the route calculation device 306 uses the current position of the vehicle acquired by the information acquisition device 301 as the departure point and the information indicated by the voice recognition result as the destination. The route from the starting point to the destination is calculated based on the map data. The route guidance device 307 outputs the route guidance information calculated according to the route calculated by the route calculation unit 306 via the output device 303, and performs route guidance for the user.

As described above, according to the fifth embodiment, the speech recognition apparatus 100 generates a speech recognition result indicating a favorable speech recognition rate for noise speech data including a user's utterance input to the input device 304. The noise suppression unit 3 that is predicted to perform the noise suppression process performs speech recognition, so that the route calculation can be performed based on the speech recognition result with a good speech recognition rate. Route guidance can be performed.

In the fifth embodiment described above, the configuration in which the voice recognition device 100 described in the first embodiment is applied to the navigation system 300 has been described. However, the voice recognition device 100a described in the second embodiment and the third embodiment are described. The speech recognition device 100b shown in FIG. 5 or the speech enhancement device 200 shown in Embodiment 4 may be applied. When the speech enhancement apparatus 200 is applied to the navigation system 300, the navigation system 300 side has a function of recognizing the enhanced speech.

In addition to the above, within the scope of the present invention, the present invention can be freely combined with each embodiment, modified any component of each embodiment, or omitted any component in each embodiment. Is possible.

Since the speech recognition apparatus and the speech enhancement apparatus according to the present invention can select a noise suppression method that provides a favorable speech recognition rate or acoustic index, a device having a call function such as a navigation system, a telephone-compatible system, and an elevator Can be applied to.

1 1st prediction unit, 1a 2nd prediction unit, 2, 2a, 2b suppression method selection unit, 3, 3a, 3b, 3c noise suppression unit, 4 speech recognition unit, 5 feature quantity calculation unit, 6, 6a similarity Degree calculation unit, 7 recognition rate database, 8 acoustic index database, 100, 100a, 100b speech recognition device, 200 speech enhancement device, 300 navigation system, 301 information acquisition device, 302 control device, 303 output device, 304 input device, 305 Map database, 306 route calculation device, 307 route guidance device.

Claims

A plurality of noise suppression units that perform noise suppression processing of different methods on the input noise voice data,
A voice recognition unit for performing voice recognition of voice data in which a noise signal is suppressed by the noise suppression unit;
A prediction unit that predicts a speech recognition rate obtained when noise suppression processing is performed on each of the noise speech data by the plurality of noise suppression units, based on an acoustic feature amount of the input noise speech data;
A speech recognition apparatus comprising: a suppression method selection unit that selects a noise suppression unit that performs noise suppression processing on the noise speech data from the plurality of noise suppression units based on the speech recognition rate predicted by the prediction unit.
The speech recognition apparatus according to claim 1, wherein the prediction unit predicts the speech recognition rate for each frame of a short-time Fourier transform of the acoustic feature amount.
The speech recognition apparatus according to claim 1, wherein the prediction unit is configured by a neural network that receives the acoustic feature amount as an input and outputs a speech recognition rate of the acoustic feature amount.
The speech according to claim 1, wherein the prediction unit includes a neural network that performs classification processing using the acoustic feature amount as an input and outputs information indicating the noise suppression unit having a high speech recognition rate. Recognition device.
The prediction unit is based on a feature amount calculation unit that calculates an acoustic feature amount in speech units from the noise voice data, and a similarity between the acoustic feature amount calculated by the feature amount calculation unit and the acoustic feature amount accumulated in advance. The speech recognition apparatus according to claim 1, further comprising a similarity calculation unit that acquires a speech recognition rate accumulated in advance.
A plurality of noise suppression units that perform noise suppression processing of different methods on the input noise voice data,
Based on the similarity between the acoustic feature amount calculated by the feature amount calculation unit and the acoustic feature amount stored in advance by the feature amount calculation unit that calculates the acoustic feature amount in units of utterances from the input noise voice data, A prediction unit having a similarity calculation unit for acquiring a pre-stored acoustic index;
A speech enhancement apparatus comprising: a suppression method selection unit that selects a noise suppression unit that performs noise suppression processing of the noise speech data from the plurality of noise suppression units based on an acoustic index acquired by the similarity calculation unit.
A predicting unit predicting a speech recognition rate obtained when noise suppression processing is performed on each of the noise speech data by the plurality of noise suppression methods, based on an acoustic feature amount of the input noise speech data;
A suppression method selection unit selecting a noise suppression unit that performs noise suppression processing on the noise speech data based on the predicted speech recognition rate;
The selected noise suppression unit performing noise suppression processing of the input noise voice data;
A speech recognition method comprising: a speech recognition unit performing speech recognition of speech data in which a noise signal is suppressed by the noise suppression processing.
A feature amount calculation unit of the prediction unit that calculates an acoustic feature amount in units of utterances from the input noise voice data;
A step of obtaining a pre-stored acoustic index based on the similarity between the calculated acoustic feature quantity and the pre-stored acoustic feature quantity, the similarity calculating section of the prediction unit;
A step of selecting a noise suppression unit that performs a noise suppression process on the noise speech data based on the acquired acoustic index;
The speech enhancement apparatus, wherein the selected noise suppression unit performs a noise suppression process on the input noise speech data.
A speech recognition device according to claim 1;
The current position of the moving object is the starting point of the moving object, the voice recognition result that is the output of the voice recognition device is the destination of the moving object, and the map data is referred to from the starting point to the destination. A route calculation device for calculating a route;
A navigation device comprising: a route guidance device that guides the movement of the mobile body according to the route calculated by the route calculation unit.